@inkeep/agents-run-api 0.40.0 → 0.41.1

package/dist/a2a/handlers.js

@@ -50,6 +50,13 @@ async function handleMessageSend(c, agent, request) {
 	try {
 		const params = request.params;
 		const { agentId } = getRequestExecutionContext(c);
+		const forwardedHeaders = {};
+		const xForwardedCookie = c.req.header("x-forwarded-cookie");
+		const authorization = c.req.header("authorization");
+		const cookie = c.req.header("cookie");
+		if (xForwardedCookie) forwardedHeaders["x-forwarded-cookie"] = xForwardedCookie;
+		else if (cookie) forwardedHeaders["x-forwarded-cookie"] = cookie;
+		if (authorization) forwardedHeaders.authorization = authorization;
 		const task = {
 			id: generateId(),
 			input: { parts: params.message.parts.map((part) => ({
@@ -62,7 +69,8 @@ async function handleMessageSend(c, agent, request) {
 				metadata: {
 					blocking: params.configuration?.blocking ?? false,
 					custom: { agent_id: agentId || "" },
-					...params.message.metadata
+					...params.message.metadata,
+					forwardedHeaders: Object.keys(forwardedHeaders).length > 0 ? forwardedHeaders : void 0
 				}
 			}
 		};
@@ -297,6 +305,13 @@ async function handleMessageStream(c, agent, request) {
 			},
 			id: request.id
 		});
+		const forwardedHeaders = {};
+		const xForwardedCookie = c.req.header("x-forwarded-cookie");
+		const authorization = c.req.header("authorization");
+		const cookie = c.req.header("cookie");
+		if (xForwardedCookie) forwardedHeaders["x-forwarded-cookie"] = xForwardedCookie;
+		else if (cookie) forwardedHeaders["x-forwarded-cookie"] = cookie;
+		if (authorization) forwardedHeaders.authorization = authorization;
 		const task = {
 			id: generateId(),
 			input: { parts: params.message.parts.map((part) => ({
@@ -308,7 +323,8 @@ async function handleMessageStream(c, agent, request) {
 				conversationId: params.message.contextId,
 				metadata: {
 					blocking: false,
-					custom: { agent_id: agentId || "" }
+					custom: { agent_id: agentId || "" },
+					forwardedHeaders: Object.keys(forwardedHeaders).length > 0 ? forwardedHeaders : void 0
 				}
 			}
 		};

package/dist/agents/Agent.d.ts

@@ -48,6 +48,8 @@ type AgentConfig = {
   sandboxConfig?: SandboxConfig;
   /** User ID for user-scoped credential lookup (from temp JWT) */
   userId?: string;
+  /** Headers to forward to MCP servers (e.g., x-forwarded-cookie for user session auth) */
+  forwardedHeaders?: Record<string, string>;
 };
 type ExternalAgentRelationConfig = {
   relationId: string;

package/dist/agents/Agent.js

@@ -5,7 +5,6 @@ import { toolSessionManager } from "./ToolSessionManager.js";
 import { getCompressionConfigForModel } from "../utils/model-context-utils.js";
 import { setSpanWithError as setSpanWithError$1, tracer } from "../utils/tracer.js";
 import { getModelAwareCompressionConfig } from "../services/BaseCompressor.js";
-import "../services/ConversationCompressor.js";
 import { createDefaultConversationHistoryConfig, getConversationHistoryWithCompression } from "../data/conversations.js";
 import { getStreamHelper } from "../utils/stream-registry.js";
 import { agentSessionManager } from "../services/AgentSession.js";
@@ -25,7 +24,7 @@ import { Phase2Config } from "./versions/v1/Phase2Config.js";
 import { z } from "@hono/zod-openapi";
 import { ContextResolver, CredentialStuffer, MCPServerType, MCPTransportType, McpClient, ModelFactory, TemplateEngine, agentHasArtifactComponents, createMessage, generateId, getContextConfigById, getCredentialReference, getFullAgentDefinition, getFunction, getFunctionToolsForSubAgent, getLedgerArtifacts, getToolsForAgent, getUserScopedCredentialReference, listTaskIdsByContextId, parseEmbeddedJson } from "@inkeep/agents-core";
 import { SpanStatusCode, trace } from "@opentelemetry/api";
-import { generateObject, generateText, streamObject, streamText, tool } from "ai";
+import { Output, generateText, streamText, tool } from "ai";
 
 //#region src/agents/Agent.ts
 /**
@@ -540,7 +539,8 @@ var Agent = class {
 		};
 	}
 	async getMcpTool(tool$1) {
-		const cacheKey = `${this.config.tenantId}-${this.config.projectId}-${tool$1.id}-${tool$1.credentialReferenceId || "no-cred"}`;
+		const forwardedHeadersHash = this.config.forwardedHeaders ? Object.keys(this.config.forwardedHeaders).sort().join(",") : "no-fwd";
+		const cacheKey = `${this.config.tenantId}-${this.config.projectId}-${tool$1.id}-${tool$1.credentialReferenceId || "no-cred"}-${forwardedHeadersHash}`;
 		const credentialReferenceId = tool$1.credentialReferenceId;
 		const toolRelation = (await getToolsForAgent(dbClient_default)({ scopes: {
 			tenantId: this.config.tenantId,
@@ -627,11 +627,16 @@ var Agent = class {
 			else urlObj.searchParams.set("user_id", `${this.config.tenantId}||${this.config.projectId}`);
 			serverConfig.url = urlObj.toString();
 		}
+		if (this.config.forwardedHeaders && Object.keys(this.config.forwardedHeaders).length > 0) serverConfig.headers = {
+			...serverConfig.headers,
+			...this.config.forwardedHeaders
+		};
 		logger.info({
 			toolName: tool$1.name,
 			credentialReferenceId,
 			transportType: serverConfig.type,
-			headers: tool$1.headers
+			headers: tool$1.headers,
+			hasForwardedHeaders: !!this.config.forwardedHeaders
 		}, "Built MCP server config with credentials");
 		let client = this.mcpClientCache.get(cacheKey);
 		if (client && !client.isConnected()) {
@@ -1042,7 +1047,7 @@ var Agent = class {
 	}
 	getArtifactTools() {
 		return tool({
-			description: "Call this tool to get the complete artifact data with the given artifactId. This retrieves the full artifact content (not just the summary). Only use this when you need the complete artifact data and the summary shown in your context is insufficient.",
+			description: "Call this tool to retrieve EXISTING artifacts that were previously created and saved. This tool is for accessing artifacts that already exist, NOT for extracting tool results. Only use this when you need the complete artifact data and the summary shown in your context is insufficient.",
 			inputSchema: z.object({
 				artifactId: z.string().describe("The unique identifier of the artifact to get."),
 				toolCallId: z.string().describe("The tool call ID associated with this artifact.")
@@ -1638,11 +1643,13 @@ ${typeof cleanResult === "string" ? cleanResult : JSON.stringify(cleanResult, nu
 				if (error && typeof error === "object" && "name" in error && error.name === "connection_refused") return true;
 			}
 		}
-		if (steps.length >= 2) {
-			const previousStep = steps[steps.length - 2];
-			if (previousStep && "toolCalls" in previousStep && previousStep.toolCalls) {
-				const stopToolNames = includeThinkingComplete ? ["transfer_to_", "thinking_complete"] : ["transfer_to_"];
-				if (previousStep.toolCalls.some((tc) => stopToolNames.some((toolName) => toolName.endsWith("_") ? tc.toolName.startsWith(toolName) : tc.toolName === toolName)) && "toolResults" in previousStep && previousStep.toolResults) return true;
+		if (steps.length >= 1) {
+			const currentStep = steps[steps.length - 1];
+			if (currentStep && "toolCalls" in currentStep && currentStep.toolCalls) {
+				const hasTransferTool = currentStep.toolCalls.some((tc) => tc.toolName.startsWith("transfer_to_"));
+				const hasThinkingComplete = currentStep.toolCalls.some((tc) => tc.toolName === "thinking_complete");
+				if (hasTransferTool) return true;
+				if (includeThinkingComplete && hasThinkingComplete && "toolResults" in currentStep && currentStep.toolResults) return true;
 			}
 		}
 		return steps.length >= this.getMaxGenerationSteps();
@@ -1772,8 +1779,13 @@ ${output}${structureHintsFormatted}`;
 			componentSchemas.push(ArtifactReferenceSchema.getSchema());
 		}
 		let dataComponentsSchema;
-		if (componentSchemas.length === 1) dataComponentsSchema = componentSchemas[0];
-		else dataComponentsSchema = z.union(componentSchemas);
+		if (componentSchemas.length === 1) {
+			dataComponentsSchema = componentSchemas[0];
+			logger.info({ agentId: this.config.id }, "Using single schema (no union needed)");
+		} else {
+			dataComponentsSchema = z.union(componentSchemas);
+			logger.info({ agentId: this.config.id }, "Created union schema");
+		}
 		return dataComponentsSchema;
 	}
 	calculatePhase2Timeout(structuredModelSettings) {
@@ -1808,15 +1820,15 @@ ${output}${structureHintsFormatted}`;
 		return phase2Messages;
 	}
 	async executeStreamingPhase2(structuredModelSettings, phase2Messages, dataComponentsSchema, phase2TimeoutMs, sessionId, contextId, response) {
-		const streamResult = streamObject({
+		const streamResult = streamText({
 			...structuredModelSettings,
 			messages: phase2Messages,
-			schema: z.object({ dataComponents: z.array(dataComponentsSchema) }),
+			output: Output.object({ schema: z.object({ dataComponents: z.array(dataComponentsSchema) }) }),
 			experimental_telemetry: this.buildTelemetryConfig("structured_generation"),
 			abortSignal: AbortSignal.timeout(phase2TimeoutMs)
 		});
 		const parser = this.setupStreamParser(sessionId, contextId);
-		for await (const delta of streamResult.partialObjectStream) if (delta) await parser.processObjectDelta(delta);
+		for await (const delta of streamResult.partialOutputStream) if (delta) await parser.processObjectDelta(delta);
 		await parser.finalize();
 		const structuredResponse = await streamResult;
 		const collectedParts = parser.getCollectedParts();
@@ -1827,22 +1839,22 @@ ${output}${structureHintsFormatted}`;
 		})) };
 		return {
 			...response,
-			object: structuredResponse.object,
-			textResponse: JSON.stringify(structuredResponse.object, null, 2)
+			object: structuredResponse.output,
+			textResponse: JSON.stringify(structuredResponse.output, null, 2)
 		};
 	}
 	async executeNonStreamingPhase2(structuredModelSettings, phase2Messages, dataComponentsSchema, phase2TimeoutMs, response) {
-		const structuredResponse = await generateObject(withJsonPostProcessing({
+		const structuredResponse = await generateText(withJsonPostProcessing({
 			...structuredModelSettings,
 			messages: phase2Messages,
-			schema: z.object({ dataComponents: z.array(dataComponentsSchema) }),
+			output: Output.object({ schema: z.object({ dataComponents: z.array(dataComponentsSchema) }) }),
 			experimental_telemetry: this.buildTelemetryConfig("structured_generation"),
 			abortSignal: AbortSignal.timeout(phase2TimeoutMs)
 		}));
 		return {
 			...response,
-			object: structuredResponse.object,
-			textResponse: JSON.stringify(structuredResponse.object, null, 2)
+			object: structuredResponse.output,
+			textResponse: JSON.stringify(structuredResponse.output, null, 2)
 		};
 	}
 	async formatFinalResponse(response, textResponse, sessionId, contextId) {

package/dist/agents/generateTaskHandler.js

@@ -21,6 +21,7 @@ const createTaskHandler = (config, credentialStoreRegistry) => {
 				},
 				artifacts: []
 			};
+			const forwardedHeaders = task.context?.metadata?.forwardedHeaders;
 			const [internalRelations, externalRelations, teamRelations, toolsForAgent, dataComponents, artifactComponents] = await Promise.all([
 				getRelatedAgentsForAgent(dbClient_default)({
 					scopes: {
@@ -322,7 +323,8 @@ const createTaskHandler = (config, credentialStoreRegistry) => {
 				artifactComponents,
 				contextConfigId: config.contextConfigId || void 0,
 				conversationHistoryConfig: config.conversationHistoryConfig,
-				sandboxConfig: config.sandboxConfig
+				sandboxConfig: config.sandboxConfig,
+				forwardedHeaders
 			}, credentialStoreRegistry);
 			const artifactStreamRequestId = task.context?.metadata?.streamRequestId;
 			if (artifactStreamRequestId && artifactComponents.length > 0) agentSessionManager.updateArtifactComponents(artifactStreamRequestId, artifactComponents);

package/dist/agents/relationTools.js

@@ -38,14 +38,24 @@ ${config.tools.map((tool$1) => {
 		const toolsList = tool$1.availableTools?.map((t) => `  - ${t.name}: ${t.description || "No description available"}`).join("\n") || "";
 		return `MCP Server: ${tool$1.name}\n${toolsList}`;
 	}).join("\n\n")}`;
-	return `Hand off the conversation to agent ${config.id}.
+	return `🚨 CRITICAL TRANSFER PROTOCOL 🚨
+
+This tool immediately transfers conversation control to agent ${config.id}. 
+
+⚠️ MANDATORY BEHAVIOR:
+1. DO NOT write any response to the user
+2. DO NOT explain what you're doing  
+3. DO NOT provide partial answers
+4. ONLY call this tool and STOP
 
 Agent Information:
 - ID: ${config.id}
 - Name: ${config.name ?? "No name provided"}
 - Description: ${config.description ?? "No description provided"}${toolsSection}${transferSection}${delegateSection}
 
-Hand off the conversation to agent ${config.id} when the user's request would be better handled by this specialized agent.`;
+🔄 Use when: The user's request is better handled by this specialized agent.
+
+⛔ VIOLATION WARNING: Any text generation before/after this tool call will create a disjointed user experience. The receiving agent will provide the complete response.`;
 };
 const generateDelegateToolDescription = (delegateRelation) => {
 	const config = delegateRelation.config;
@@ -70,7 +80,7 @@ ${agentConfig.transferRelations.map((transfer) => `  - ${transfer.name || transf
 Can Delegate To:
 ${agentConfig.delegateRelations.map((delegate) => `  - ${delegate.config.name || delegate.config.id}: ${delegate.config.description || "No description available"} (${delegate.type})`).join("\n")}`;
 	}
-	return `Delegate a specific task to another agent.
+	return `Delegate a specific task to another agent and wait for their response.
 
 Agent Information:
 - ID: ${config.id}
@@ -78,7 +88,9 @@ Agent Information:
 - Description: ${config.description || "No description provided"}
 - Type: ${delegateRelation.type}${toolsSection}${transferSection}${delegateSection}
 
-Delegate a specific task to agent ${config.id} when it seems like the agent can do relevant work.`;
+Delegate a specific task to agent ${config.id} when it can do relevant work. The delegated agent will return results that you can incorporate into your response to the user.
+
+NOTE: Unlike transfers, delegation returns control back to you with the delegated agent's results.`;
 };
 const createTransferToAgentTool = ({ transferConfig, callingAgentId, subAgent, streamRequestId }) => {
 	return tool({

package/dist/agents/versions/v1/Phase1Config.js

@@ -103,9 +103,20 @@ var Phase1Config = class Phase1Config {
 		if (!hasTransferRelations) return "";
 		return `You are part of a single unified assistant composed of specialized agents. To the user, you must always appear as one continuous, confident voice.
 
-You have transfer_to_* tools that seamlessly continue the conversation. When you determine another agent should handle a request: ONLY call the appropriate transfer_to_* tool. Do not provide any substantive answer, limitation, or explanation before transferring. NEVER announce, describe, or apologize for a transfer.
-
-Do NOT stream any text when transferring - call the transfer tool IMMEDIATELY. Do NOT acknowledge the request, do NOT say "Looking into that...", "Let me search...", "I'll help you find...", or provide ANY explanatory text. Place all reasoning or handoff details inside the transfer tool call, not in the user message. The tool call is sufficient - no additional text should be generated.
+🚨 CRITICAL TRANSFER PROTOCOL 🚨
+When you determine another agent should handle a request:
+1. IMMEDIATELY call the appropriate transfer_to_* tool  
+2. Generate ZERO text in your response - no words, no explanations, no acknowledgments
+3. Do NOT stream any content - the tool call must be your ONLY output
+
+FORBIDDEN BEFORE TRANSFERS:
+❌ Do NOT acknowledge the request ("I understand you want...")
+❌ Do NOT provide partial answers ("The basics are..." then transfer) 
+❌ Do NOT explain what you're doing ("Let me search...", "I'll help you find...")
+❌ Do NOT apologize or announce transfers ("I'll need to transfer you...")
+❌ Do NOT generate ANY text content whatsoever - just call the transfer tool
+
+REMEMBER: Tool call = complete response. No additional text generation allowed.
 
 CRITICAL: When you receive a user message that ends with "Please continue from where this conversation was left off" - this indicates you are continuing a conversation that another agent started. You should:
 - Review the conversation history to see what was already communicated to the user

package/dist/create-app.d.ts

@@ -1,9 +1,9 @@
 import { SandboxConfig } from "./types/execution-context.js";
 import { CredentialStoreRegistry, ServerConfig } from "@inkeep/agents-core";
 import { Hono } from "hono";
-import * as hono_types0 from "hono/types";
+import * as hono_types3 from "hono/types";
 
 //#region src/create-app.d.ts
-declare function createExecutionHono(serverConfig: ServerConfig, credentialStores: CredentialStoreRegistry, sandboxConfig?: SandboxConfig): Hono<hono_types0.BlankEnv, hono_types0.BlankSchema, "/">;
+declare function createExecutionHono(serverConfig: ServerConfig, credentialStores: CredentialStoreRegistry, sandboxConfig?: SandboxConfig): Hono<hono_types3.BlankEnv, hono_types3.BlankSchema, "/">;
 //#endregion
 export { createExecutionHono };

package/dist/handlers/executionHandler.d.ts

@@ -10,6 +10,8 @@ interface ExecutionHandlerParams {
   requestId: string;
   sseHelper: StreamHelper;
   emitOperations?: boolean;
+  /** Headers to forward to MCP servers (e.g., x-forwarded-cookie for auth) */
+  forwardedHeaders?: Record<string, string>;
 }
 interface ExecutionResult {
   success: boolean;

package/dist/handlers/executionHandler.js

@@ -32,7 +32,7 @@ var ExecutionHandler = class {
 	* @returns
 	*/
 	async execute(params) {
-		const { executionContext, conversationId, userMessage, initialAgentId, requestId, sseHelper, emitOperations } = params;
+		const { executionContext, conversationId, userMessage, initialAgentId, requestId, sseHelper, emitOperations, forwardedHeaders } = params;
 		const { tenantId, projectId, agentId, apiKey, baseUrl } = executionContext;
 		registerStreamHelper(requestId, sseHelper);
 		agentSessionManager.createSession(requestId, agentId, tenantId, projectId, conversationId);
@@ -177,7 +177,8 @@ var ExecutionHandler = class {
 					"x-inkeep-tenant-id": tenantId,
 					"x-inkeep-project-id": projectId,
 					"x-inkeep-agent-id": agentId,
-					"x-inkeep-sub-agent-id": currentAgentId
+					"x-inkeep-sub-agent-id": currentAgentId,
+					...forwardedHeaders || {}
 				} });
 				let messageResponse = null;
 				const messageMetadata = { stream_request_id: requestId };

package/dist/index.d.ts

@@ -3,14 +3,14 @@ import { createExecutionHono } from "./create-app.js";
 import "./env.js";
 import { CredentialStore, ServerConfig } from "@inkeep/agents-core";
 import { Hono } from "hono";
-import * as hono_types1 from "hono/types";
+import * as hono_types0 from "hono/types";
 
 //#region src/index.d.ts
-declare const app: Hono<hono_types1.BlankEnv, hono_types1.BlankSchema, "/">;
+declare const app: Hono<hono_types0.BlankEnv, hono_types0.BlankSchema, "/">;
 declare function createExecutionApp(config?: {
   serverConfig?: ServerConfig;
   credentialStores?: CredentialStore[];
   sandboxConfig?: SandboxConfig;
-}): Hono<hono_types1.BlankEnv, hono_types1.BlankSchema, "/">;
+}): Hono<hono_types0.BlankEnv, hono_types0.BlankSchema, "/">;
 //#endregion
 export { Hono, type NativeSandboxConfig, type SandboxConfig, type VercelSandboxConfig, createExecutionApp, createExecutionHono, app as default };

package/dist/routes/chat.js

@@ -108,12 +108,18 @@ app.openapi(chatCompletionsRoute, async (c) => {
 		}, "Extracted chat parameters from API key context");
 		const body = c.get("requestBody") || {};
 		const conversationId = body.conversationId || getConversationId();
+		const targetTenantId = c.req.header("x-target-tenant-id");
+		const targetProjectId = c.req.header("x-target-project-id");
+		const targetAgentId = c.req.header("x-target-agent-id");
 		const activeSpan = trace.getActiveSpan();
 		if (activeSpan) activeSpan.setAttributes({
 			"conversation.id": conversationId,
 			"tenant.id": tenantId,
 			"agent.id": agentId,
-			"project.id": projectId
+			"project.id": projectId,
+			...targetTenantId && { "target.tenant.id": targetTenantId },
+			...targetProjectId && { "target.project.id": targetProjectId },
+			...targetAgentId && { "target.agent.id": targetAgentId }
 		});
 		let currentBag = propagation.getBaggage(context.active());
 		if (!currentBag) currentBag = propagation.createBaggage();
@@ -244,6 +250,11 @@ app.openapi(chatCompletionsRoute, async (c) => {
 					await sseHelper.writeRole();
 					logger.info({ subAgentId }, "Starting execution");
 					const emitOperations = c.req.header("x-emit-operations") === "true";
+					const forwardedHeaders = {};
+					const xForwardedCookie = c.req.header("x-forwarded-cookie");
+					const cookie = c.req.header("cookie");
+					if (xForwardedCookie) forwardedHeaders["x-forwarded-cookie"] = xForwardedCookie;
+					else if (cookie) forwardedHeaders["x-forwarded-cookie"] = cookie;
 					const result = await new ExecutionHandler().execute({
 						executionContext,
 						conversationId,
@@ -251,7 +262,8 @@ app.openapi(chatCompletionsRoute, async (c) => {
 						initialAgentId: subAgentId,
 						requestId,
 						sseHelper,
-						emitOperations
+						emitOperations,
+						forwardedHeaders
 					});
 					logger.info({ result }, `Execution completed: ${result.success ? "success" : "failed"} after ${result.iterations} iterations`);
 					if (!result.success) await sseHelper.writeOperation(errorOp("Sorry, I was unable to process your request at this time. Please try again.", "system"));

package/dist/routes/chatDataStream.js

@@ -72,12 +72,23 @@ app.openapi(chatDataStreamRoute, async (c) => {
 		}, "Extracted chatDataStream parameters");
 		const body = c.get("requestBody") || {};
 		const conversationId = body.conversationId || getConversationId();
+		const targetTenantId = c.req.header("x-target-tenant-id");
+		const targetProjectId = c.req.header("x-target-project-id");
+		const targetAgentId = c.req.header("x-target-agent-id");
+		const forwardedHeaders = {};
+		const xForwardedCookie = c.req.header("x-forwarded-cookie");
+		const cookie = c.req.header("cookie");
+		if (xForwardedCookie) forwardedHeaders["x-forwarded-cookie"] = xForwardedCookie;
+		else if (cookie) forwardedHeaders["x-forwarded-cookie"] = cookie;
 		const activeSpan = trace.getActiveSpan();
 		if (activeSpan) activeSpan.setAttributes({
 			"conversation.id": conversationId,
 			"tenant.id": tenantId,
 			"agent.id": agentId,
-			"project.id": projectId
+			"project.id": projectId,
+			...targetTenantId && { "target.tenant.id": targetTenantId },
+			...targetProjectId && { "target.project.id": targetProjectId },
+			...targetAgentId && { "target.agent.id": targetAgentId }
 		});
 		let currentBag = propagation.getBaggage(context.active());
 		if (!currentBag) currentBag = propagation.createBaggage();
@@ -177,7 +188,8 @@ app.openapi(chatDataStreamRoute, async (c) => {
 					initialAgentId: subAgentId,
 					requestId: `chat-${Date.now()}`,
 					sseHelper: bufferingHelper,
-					emitOperations
+					emitOperations,
+					forwardedHeaders
 				});
 				const captured = bufferingHelper.getCapturedResponse();
 				return c.json({
@@ -211,7 +223,8 @@ app.openapi(chatDataStreamRoute, async (c) => {
 						initialAgentId: subAgentId,
 						requestId: `chatds-${Date.now()}`,
 						sseHelper: streamHelper,
-						emitOperations
+						emitOperations,
+						forwardedHeaders
 					})).success) await streamHelper.writeOperation(errorOp("Unable to process request", "system"));
 				} catch (err) {
 					logger.error({ err }, "Streaming error");

package/dist/services/AgentSession.js

@@ -11,7 +11,7 @@ import { ArtifactParser } from "./ArtifactParser.js";
 import { z } from "@hono/zod-openapi";
 import { CONVERSATION_HISTORY_DEFAULT_LIMIT, CONVERSATION_HISTORY_MAX_OUTPUT_TOKENS_DEFAULT, ModelFactory, getLedgerArtifacts, getSubAgentById } from "@inkeep/agents-core";
 import { SpanStatusCode } from "@opentelemetry/api";
-import { generateObject } from "ai";
+import { Output, generateText } from "ai";
 
 //#region src/services/AgentSession.ts
 const logger = getLogger("AgentSession");
@@ -517,16 +517,22 @@ var AgentSession = class {
 					}, "Failed to fetch conversation history for structured status update");
 				}
 				const previousSummaryContext = previousSummaries.length > 0 ? `\nPrevious updates sent to user:\n${previousSummaries.map((s, i) => `${i + 1}. ${s}`).join("\n")}\n` : "";
-				const selectionSchema = z.object(Object.fromEntries([["no_relevant_updates", z.object({ no_updates: z.boolean().default(true) }).optional().describe("Use when nothing substantially new to report. Should only use on its own.")], ...statusComponents.map((component) => [component.type, this.getComponentSchema(component).optional().describe(component.description || component.type)])]));
+				const selectionSchema = z.object({ updates: z.array(z.union([z.object({
+					type: z.literal("no_relevant_updates"),
+					data: z.object({ no_updates: z.boolean().default(true) }).describe("Use when nothing substantially new to report. Should only use on its own.")
+				}), ...statusComponents.map((component) => z.object({
+					type: z.literal(component.type),
+					data: this.getComponentSchema(component).describe(component.description || component.type)
+				}))])) });
 				const prompt = `Generate status updates for relevant components based on what the user has asked for.${conversationContext}${previousSummaries.length > 0 ? `\n${previousSummaryContext}` : ""}
 
 Activities:\n${userVisibleActivities.join("\n") || "No New Activities"}
 
-Available components: no_relevant_updates, ${statusComponents.map((c) => c.type).join(", ")}
+Available component types: no_relevant_updates, ${statusComponents.map((c) => c.type).join(", ")}
 
 Rules:
-- Fill in data for relevant components only
-- Use 'no_relevant_updates' if nothing substantially new to report. DO NOT WRITE LABELS OR USE OTHER COMPONENTS IF YOU USE THIS COMPONENT.
+- Return an array of updates for relevant components
+- Use 'no_relevant_updates' type if nothing substantially new to report. DO NOT INCLUDE OTHER COMPONENT TYPES IF YOU USE THIS ONE.
 - Never repeat previous values, make every update EXTREMELY unique. If you cannot do that the update is not worth mentioning.
 - Labels MUST be short 3-7 word phrases with ACTUAL information discovered. NEVER MAKE UP SOMETHING WITHOUT BACKING IT UP WITH ACTUAL INFORMATION.
 - Use sentence case: only capitalize the first word and proper nouns (e.g., "Admin permissions required", not "Admin Permissions Required"). ALWAYS capitalize the first word of the label.
@@ -577,10 +583,10 @@ ${this.statusUpdateState?.config.prompt?.trim() || ""}`;
 					modelToUse = this.statusUpdateState.baseModel;
 				}
 				if (!modelToUse) throw new Error("No model configuration available");
-				const { object } = await generateObject({
+				const { output: object } = await generateText({
 					model: ModelFactory.createModel(modelToUse),
 					prompt,
-					schema: selectionSchema,
+					output: Output.object({ schema: selectionSchema }),
 					experimental_telemetry: {
 						isEnabled: true,
 						functionId: `structured_update_${this.sessionId}`,
@@ -595,21 +601,19 @@ ${this.statusUpdateState?.config.prompt?.trim() || ""}`;
 				const result = object;
 				logger.info({ result: JSON.stringify(result) }, "DEBUG: Result");
 				const summaries = [];
-				for (const [componentId, data] of Object.entries(result)) {
-					logger.info({
-						componentId,
-						data: JSON.stringify(data)
-					}, "DEBUG: Component data");
-					if (componentId === "no_relevant_updates") continue;
-					if (data && typeof data === "object" && Object.keys(data).length > 0) summaries.push({
-						type: componentId,
-						data
+				const updates = result.updates || [];
+				for (const update of updates) {
+					logger.info({ update: JSON.stringify(update) }, "DEBUG: Update data");
+					if (update.type === "no_relevant_updates") continue;
+					if (update.data && typeof update.data === "object" && Object.keys(update.data).length > 0) summaries.push({
+						type: update.type,
+						data: update.data
 					});
 				}
 				span.setAttributes({
 					"summaries.count": summaries.length,
 					"user_activities.count": userVisibleActivities.length,
-					"result_keys.count": Object.keys(result).length
+					"updates.count": updates.length
 				});
 				span.setStatus({ code: SpanStatusCode.OK });
 				return { summaries };
@@ -875,7 +879,7 @@ Make the name extremely specific to what this tool call actually returned, not g
 						name: z.string().describe("Concise, descriptive name for the artifact"),
 						description: z.string().describe("Brief description of the artifact's relevance to the user's question")
 					});
-					const { object } = await tracer.startActiveSpan("agent_session.generate_artifact_metadata", { attributes: {
+					const { output: object } = await tracer.startActiveSpan("agent_session.generate_artifact_metadata", { attributes: {
 						"llm.model": this.statusUpdateState?.summarizerModel?.model,
 						"llm.operation": "generate_object",
 						"artifact.id": artifactData.artifactId,
@@ -887,10 +891,10 @@ Make the name extremely specific to what this tool call actually returned, not g
 						const maxRetries = 3;
 						let lastError = null;
 						for (let attempt = 1; attempt <= maxRetries; attempt++) try {
-							const result$1 = await generateObject({
+							const result$1 = await generateText({
 								model,
 								prompt,
-								schema,
+								output: Output.object({ schema }),
 								experimental_telemetry: {
 									isEnabled: true,
 									functionId: `artifact_processing_${artifactData.artifactId}`,
@@ -906,12 +910,12 @@ Make the name extremely specific to what this tool call actually returned, not g
 							generationSpan.setAttributes({
 								"artifact.id": artifactData.artifactId,
 								"artifact.type": artifactData.artifactType,
-								"artifact.name": result$1.object.name,
-								"artifact.description": result$1.object.description,
+								"artifact.name": result$1.output.name,
+								"artifact.description": result$1.output.description,
 								"artifact.summary": JSON.stringify(artifactData.summaryData, null, 2),
 								"artifact.full": JSON.stringify(artifactData.data || artifactData.summaryData, null, 2),
-								"generation.name_length": result$1.object.name.length,
-								"generation.description_length": result$1.object.description.length,
+								"generation.name_length": result$1.output.name.length,
+								"generation.description_length": result$1.output.description.length,
 								"generation.attempts": attempt
 							});
 							generationSpan.setStatus({ code: SpanStatusCode.OK });

package/dist/services/BaseCompressor.js

@@ -409,7 +409,6 @@ var BaseCompressor = class {
 		return await tracer.startActiveSpan("compressor.safe_compress", { attributes: {
 			"compression.type": this.getCompressionType(),
 			"compression.session_id": this.sessionId,
-			"compression.conversation_id": this.conversationId,
 			"compression.message_count": messages.length,
 			"compression.input_tokens": fullContextSize ?? this.calculateContextSize(messages),
 			"compression.hard_limit": this.getHardLimit(),
@@ -423,7 +422,7 @@ var BaseCompressor = class {
 					"compression.result.output_tokens": resultTokens,
 					"compression.result.compression_ratio": (fullContextSize ?? this.calculateContextSize(messages)) > 0 ? ((fullContextSize ?? this.calculateContextSize(messages)) - resultTokens) / (fullContextSize ?? this.calculateContextSize(messages)) : 0,
 					"compression.success": true,
-					"compression.fallback_used": false
+					"compression.result.summary": result.summary?.high_level || ""
 				});
 				compressionSpan.setStatus({ code: SpanStatusCode.OK });
 				return result;
@@ -434,10 +433,7 @@ var BaseCompressor = class {
 					error: error instanceof Error ? error.message : String(error),
 					stack: error instanceof Error ? error.stack : void 0
 				}, "Compression failed, using simple fallback");
-				compressionSpan.setAttributes({
-					"compression.error": error instanceof Error ? error.message : String(error),
-					"compression.fallback_used": true
-				});
+				compressionSpan.setAttributes({ "compression.error": error instanceof Error ? error.message : String(error) });
 				const fallbackResult = await this.simpleCompressionFallback(messages);
 				const fallbackTokens = Array.isArray(fallbackResult.summary) ? this.calculateContextSize(fallbackResult.summary) : this.estimateTokens(fallbackResult.summary);
 				compressionSpan.setAttributes({

package/dist/tools/distill-conversation-history-tool.js

@@ -1,6 +1,6 @@
 import { getLogger } from "../logger.js";
 import { ModelFactory } from "@inkeep/agents-core";
-import { generateObject } from "ai";
+import { Output, generateText } from "ai";
 import { z } from "zod";
 
 //#region src/tools/distill-conversation-history-tool.ts
@@ -57,7 +57,7 @@ async function distillConversationHistory(params) {
 	const { messages, conversationId, summarizerModel, toolCallToArtifactMap } = params;
 	try {
 		if (!summarizerModel?.model?.trim()) throw new Error("Summarizer model is required");
-		const { object: summary } = await generateObject({
+		const { output: summary } = await generateText({
 			model: ModelFactory.createModel(summarizerModel),
 			prompt: `You are a conversation history summarization assistant. Your job is to create a comprehensive summary that can COMPLETELY REPLACE the original conversation history while preserving all essential context.
 
@@ -158,7 +158,7 @@ Create a comprehensive summary using this exact JSON schema:
 **REMEMBER**: This summary is REPLACING the entire conversation history. Include everything essential for context continuation.
 
 Return **only** valid JSON.`,
-			schema: ConversationHistorySummarySchema
+			output: Output.object({ schema: ConversationHistorySummarySchema })
 		});
 		summary.session_id = conversationId;
 		return summary;

package/dist/tools/distill-conversation-tool.js

@@ -1,6 +1,6 @@
 import { getLogger } from "../logger.js";
 import { ModelFactory } from "@inkeep/agents-core";
-import { generateObject, tool } from "ai";
+import { Output, generateText } from "ai";
 import { z } from "zod";
 
 //#region src/tools/distill-conversation-tool.ts
@@ -36,7 +36,7 @@ async function distillConversation(params) {
 	try {
 		const modelToUse = summarizerModel;
 		if (!modelToUse?.model?.trim()) throw new Error("Summarizer model is required");
-		const { object: summary } = await generateObject({
+		const { output: summary } = await generateText({
 			model: ModelFactory.createModel(modelToUse),
 			prompt: `You are a conversation summarization assistant. Your job is to create or update a compact, structured summary that captures VALUABLE CONTENT and FINDINGS, not just operational details.
 
@@ -111,7 +111,7 @@ Create/update a summary using this exact JSON schema:
 **Focus on WHAT WAS LEARNED, not HOW IT WAS LEARNED**
 
 Return **only** valid JSON.`,
-			schema: ConversationSummarySchema
+			output: Output.object({ schema: ConversationSummarySchema })
 		});
 		summary.session_id = conversationId;
 		return summary;

package/dist/utils/json-postprocessor.d.ts

@@ -4,7 +4,7 @@
  */
 declare function stripJsonCodeBlocks(text: string): string;
 /**
- * Configuration helper to add JSON post-processing to generateObject calls
+ * Configuration helper to add JSON post-processing to structured output generateText calls
  */
 declare function withJsonPostProcessing<T extends Record<string, any>>(config: T): T & {
   experimental_transform?: (text: string) => string;

package/dist/utils/json-postprocessor.js

@@ -6,7 +6,7 @@ function stripJsonCodeBlocks(text) {
 	return text.trim().replace(/^```json\s*/is, "").replace(/^```\s*/s, "").replace(/\s*```$/s, "").replace(/^```json\s*([\s\S]*?)\s*```$/i, "$1").replace(/^```\s*([\s\S]*?)\s*```$/i, "$1").trim();
 }
 /**
-* Configuration helper to add JSON post-processing to generateObject calls
+* Configuration helper to add JSON post-processing to structured output generateText calls
 */
 function withJsonPostProcessing(config) {
 	return {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@inkeep/agents-run-api",
-  "version": "0.40.0",
+  "version": "0.41.1",
   "description": "Agents Run API for Inkeep Agent Framework - handles chat, agent execution, and streaming",
   "types": "dist/index.d.ts",
   "exports": {
@@ -13,13 +13,13 @@
   "type": "module",
   "license": "SEE LICENSE IN LICENSE.md",
   "dependencies": {
-    "@electric-sql/pglite": "^0.3.13",
-    "@ai-sdk/anthropic": "3.0.0-beta.66",
-    "@ai-sdk/gateway": "2.0.0-beta.68",
-    "@ai-sdk/google": "3.0.0-beta.62",
-    "@ai-sdk/openai": "3.0.0-beta.74",
-    "@ai-sdk/openai-compatible": "2.0.0-beta.41",
+    "@ai-sdk/anthropic": "3.0.7",
+    "@ai-sdk/gateway": "3.0.9",
+    "@ai-sdk/google": "3.0.4",
+    "@ai-sdk/openai": "3.0.7",
+    "@ai-sdk/openai-compatible": "2.0.4",
     "@alcyone-labs/modelcontextprotocol-sdk": "^1.16.0",
+    "@electric-sql/pglite": "^0.3.13",
     "@hono/otel": "^0.4.0",
     "@hono/swagger-ui": "^0.5.1",
     "@openrouter/ai-sdk-provider": "^1.2.0",
@@ -34,14 +34,14 @@
     "@opentelemetry/sdk-trace-base": "^2.1.0",
     "@opentelemetry/semantic-conventions": "^1.37.0",
     "@vercel/sandbox": "^0.0.24",
-    "ai": "6.0.0-beta.124",
+    "ai": "6.0.14",
     "ajv": "^8.17.1",
     "drizzle-orm": "^0.44.4",
     "fetch-to-node": "^2.1.0",
     "hono": "^4.10.4",
     "jmespath": "^0.16.0",
     "llm-info": "^1.0.69",
-    "@inkeep/agents-core": "^0.40.0"
+    "@inkeep/agents-core": "^0.41.1"
   },
   "peerDependencies": {
     "@hono/zod-openapi": "^1.1.5",